Abstract:

Physics-based simulation models are essential for the development of vapor compression cycles (VCCs), the core technology behind most modern refrigeration and air conditioning systems. These models enable robust control, monitoring, fault detection and diagnostics, and digital twin technologies. However, nonlinear dynamics, high-dimensional parameter and state spaces, numerical stiffness, and the limited integration of conventional modeling environments with scientific machine learning workflows create significant challenges for efficient, transferable, and automated calibration. Existing approaches typically rely on deterministic methods and lack mechanisms for principled knowledge transfer across calibration tasks, while also failing to quantify epistemic and aleatoric uncertainty. To address these limitations, we propose a Bayesian calibration framework for VCC systems that explicitly quantifies various sources of uncertainty in model predictions. The framework supports transferability across datasets by sequentially up- dating informative priors from previously inferred posteriors. We implement and evaluate this approach on a commercially available high-fidelity Julia based dynamic VCC model, demonstrating its ability to successfully estimate key system parameters.