Comparative machine-learning modelling of effective thermal conductivity in nano-enhanced phase change materials

Organic phase change materials store latent heat efficiently but their low thermal conductivity limits charging and discharging rates. This study evaluates whether a data-driven surrogate can predict conductivity enhancement across a heterogeneous, multi-source experimental database. A curated set of 150 measurements from 21 primary studies was modelled using linear regression, support-vector regression, random forest, gradient boosting and a feed-forward neural network. The target was the dimensionless logarithmic enhancement ratio y = log 10 (k eff /k base ), evaluated with preprocessing-leakage-free ten-fold cross-validation. Gradient boosting and random forest gave the highest mean cross-validation R 2 values, 0.46 ± 0.56 and 0.43 ± 0.70, respectively; the lowest RMSE was 0.17 and the lowest MAE was 0.10. Support-vector regression and the neural network overfitted, with negative cross-validation R 2 despite high training scores. Loading and filler intrinsic conductivity were the dominant descriptors. An operational synergy index applied only to matched mono- and hybrid measurements ranged from 0.65 to 0.75 for sub-additive pairs to 2.0 for a GNP-MWCNT formulation. Leave-one-source-out validation reduced R 2 to approximately 0.05, showing that the model is suitable for screening within represented material families but not for quantitative prediction of unseen source systems without retraining.