Interpretable machine learning for COP prediction in vapour compression cycles with low-GWP refrigerants: A SHAP-based analysis of NIST experimental data

Selecting low-global-warming-potential (GWP) refrigerants requires performance estimates for fluids and operating conditions that experimental campaigns can only sparsely cover. This study presents an interpretable XGBoost surrogate framework that predicts the cooling coefficient of performance (COP) of vapour compression cycles, trained exclusively on 248 steady-state experimental observations spanning ten working fluids (NIST Technical Note 2233). The defining feature of the framework is its physics-based fluid representation: each refrigerant—pure or blended—is described by four continuous mixture descriptors (critical temperature, critical pressure, acentric factor, and molar mass) instead of categorical identifiers, embedding all fluids in a shared thermodynamic coordinate space and enabling the assessment of unseen candidates without retraining or re-encoding. The model achieved R 2 = 0.979 (RMSE = 0.144 , MAPE = 2.0 % ) on a stratified held-out test set, R 2 = 0.981 ± 0.007 under 5-fold cross-validation, and a mean R 2 of 0.896 ± 0.084 under Leave-One-Refrigerant-Out cross-validation, in which each fluid is predicted by a model that has never seen it. SHapley Additive exPlanations (SHAP) confirmed a physically consistent feature hierarchy—condenser and evaporator outlet temperatures dominate, followed by mixture critical temperature—and revealed family-resolved descriptor fingerprints (intra-family cosine similarity 0.91 versus inter-family − 0.49 ). Conformally calibrated prediction intervals (90% nominal, 87.5% empirical coverage) and an explicit applicability domain ( T c,mix ∈ [345, 382] K, P c,mix ∈ [3.4, 5.8] MPa) delimit where screening is reliable. The framework thereby offers a rapid, physically grounded, and uncertainty-aware tool for low-GWP refrigerant screening, supporting the transition toward sustainable refrigeration technologies.