An end-to-end differentiable transient vapor-compression framework for automated machine sizing and unified optimal control

The modern world runs on heat. From the warmth that keeps our homes comfortable in winter to the cool air that makes summer bearable, we rely on machines that move thermal energy from one place to another. These machines, known as heat pumps, are the unsung heroes of the global effort to switch from burning fossil fuels to using electricity. However, designing them has long been a fragmented process. Engineers typically start with a static snapshot of what a machine should do, then manually translate that into a complex computer model to see how it behaves when conditions change. Finally, they try to figure out how to control the machine, often using simplified versions of the model that no longer match the real physics. This disconnect between design, simulation, and control creates a gap where efficiency is lost and performance is harder to predict.

A researcher has now built a new kind of digital framework that bridges these gaps. Instead of treating design, simulation, and control as separate steps, they created a single, unified system that connects them all. This system is designed to automatically figure out the exact size and shape of a heat pump based on the specific job it needs to do, and then immediately use that same detailed model to control the machine in real time. By doing this, the researcher has shown that it is possible to create a machine that is perfectly matched to its task and can respond instantly to changing weather or energy grid demands, all without the errors that usually creep in when switching between different computer models.

The core of this work is a new way of describing how a heat pump works. Traditionally, these machines are modeled by breaking them down into small sections and calculating how pressure and temperature change as refrigerant flows through them. This is difficult because the refrigerant changes from a gas to a liquid and back again, and the equations that describe these changes are notoriously stubborn and hard for computers to solve quickly. The researcher solved this by pre-calculating the behavior of the refrigerant and storing it in a way that the computer can read instantly, without having to stop and solve complex math problems every time the simulation runs. This allows the entire system to be "differentiable," meaning the computer can instantly see how a tiny change in one part of the machine affects the whole system. This capability is crucial for both designing the machine and for teaching it how to control itself efficiently.

Using this framework, the researcher demonstrated an automated process that starts with a simple request: a specific amount of heating or cooling needed. The system then works backward to determine the exact physical dimensions of the machine. It calculates the size of the compressor, the area of the valve that controls the flow of refrigerant, and the number of tubes needed in the heat exchangers. In a standard example, the system designed a unit capable of providing 5.5 kilowatts of heating power. It determined that the compressor needed to displace 25.9 cubic centimeters of fluid per revolution, the valve needed a maximum opening area of 1.20 square millimeters, and the indoor and outdoor coils required 67 and 63 tubes, respectively. The system did this without human intervention, ensuring that the final design was mathematically consistent with the laws of thermodynamics.

Once the machine was designed, the researcher tested how well it performed in simulated real-world conditions. They ran the model through various scenarios, including heating a room when it was freezing outside and cooling a room on a hot day. In one test, the system successfully maintained a room temperature of 20 degrees Celsius when the outside air was at 0 degrees, with the compressor running at a steady speed. In another test, the machine switched from cooling to heating mode, and the system tracked the changes in pressure and temperature smoothly. The results showed that the automated design could predict the machine's cooling capacity with an average error of less than 8 percent when compared to real-world data from existing mini-split units. When tested against data from a large-scale laboratory experiment, the error in predicting cooling capacity was even smaller, staying within a range of 1.2 to 1.6 percent.

The study also highlighted the importance of using the same detailed model for both designing the machine and controlling it. In the past, engineers often had to simplify the model for the controller, which could lead to a mismatch where the machine behaved differently than the controller expected. By using the same complex, high-fidelity model for both tasks, the researcher eliminated this mismatch. They showed that a controller based on this unified model could track temperature setpoints accurately and respond to changes in the environment without the lag or instability that often plagues older systems. This approach allows for a level of precision that was previously difficult to achieve, as the controller can anticipate how the machine will react to its own actions.

The researcher was careful to note that their system is a simulation tool and not a physical machine. The results presented are based on computer models that have been validated against real-world data, but they are not a replacement for physical testing. The framework does not include every possible feature, such as automatic defrost cycles or complex ductwork, and it assumes the refrigerant behaves in a specific, idealized way. However, the fact that the system could predict performance so accurately without being "tuned" to fit the data suggests that the underlying physics are being captured correctly. The system is designed to be open-source, allowing other engineers to use it as a foundation for their own designs and control strategies.

This work represents a significant step forward in how thermal systems are engineered. By unifying the design and control processes into a single, mathematically consistent framework, the researcher has removed the barriers that often lead to inefficiency and error. The ability to automatically size a machine based on its intended duty and then immediately apply a sophisticated control strategy to it offers a new path for developing heat pumps that are more efficient, more responsive, and better suited to the demands of a modern, electrified grid. The findings suggest that the future of thermal management lies not in building better individual components, but in creating systems where the design and the control are inextricably linked, working together as a single, intelligent whole.