Virtual sensors are mathematical models that predict the readings of a sensor in a location currently without an operational sensor. Virtual sensors can be used to compensate for a failed sensor or as a framework for supporting mathematical decomposition of a model of a complex system. This study applies a novel genetic programming representation called a function stack to the problem of virtual sensor induction in a simple thermal system. Real-valued function stacks are introduced in this study. The thermal system modeled is a heat exchanger. Function stacks are found to be able to efficiently find compact and accurate models for each often sensors using the data from the other sensors. This study serves as proof-of-concept for using function stacks as a modeling technology for virtual sensors.