DIGITAL TWINS OF THERMAL SYSTEMS: A COMPARISON BETWEEN SUPERVISED AND REINFORCEMENT LEARNING

This article explores two novel approaches for controlling heat transfer systems through the development of “digital twins,” focusing on transient thermal systems. The study involves creating a digital representation of a physical system, specifically a 2D square subject to an inward and transient heat flux, with the goal of keeping the maximum temperature within a predefined limit by changing the convective cooling. The first method utilizes a neural network trained on steady-state data, whereas the second employs an interactive learning algorithm. Results show that both strategies prove to be effective in managing the system’s thermal performance. However, the RL-based approach demonstrates greater flexibility in adapting to new scenarios, albeit at the cost of increased computational demands due to the necessity of integrating interactive learning with unsteady finite element method (FEM) simulations for training, validation, and testing phases.