AI for dynamic heat exchangers modeling in vapor compression systems with limited computational resources

AI for Science has been regarded as the fifth paradigm of scientific research, yet the application in vapor compression system modeling remains limited. Dynamic heat exchanger models are governed by complex partial differential equations and suffer from slow computation and poor robustness. In this work, an AI-driven two-phase heat exchanger model using recurrent neural networks was proposed and a virtual-valve mechanism was introduced to enable flexible modular connection of heat exchanger modules. Based on analogy to automatic control theory, performance evaluation metrics were proposed covering accuracy, stability, and response rapidity. Although the AI-driven model exhibited slower dynamic response than physics-driven model due to the larger time steps, the AI-driven model achieved a prediction deviation below 1% while delivering a fivefold speedup in computation. Using a heat pump water heater system as an experimental demonstration, under complex boundary conditions, the AI-driven model deviated by only 5%. To optimize the training data generation strategy, orthogonal design-based dataset simplification was proposed without performance degradation. Furthermore, the trained AI-driven model maintained robust performance when deployed in a desktop computer with limited computational resources, whereas the performance of physics-driven model decreased severely. Building on these findings, this study envisioned the future development pathways for AI-driven modeling in engineering applications.