<p>Solar air heaters (SAHs) are an environmentally friendly way to heat air at lower and medium temperatures, but their efficiency is limited by the air’s low thermal conductivity. This study experimentally investigates the performance of an evacuated tube solar air heater (ETSAH) comprising 20 evacuated tubes fitted with a baffle partition and a pierced twisted tape (PTT) within each tube to enhance heat transfer. Three machine learning (ML) models, namely k-nearest neighbours (KNN), random forest (RF), and linear regression (LR), are systematically compared to predict the useful heat gain, thermal efficiency, effective efficiency, and exergy efficiency. The results indicate that LR consistently outperforms both RF and KNN, demonstrating excellent agreement between simulation and experimental data for useful heat gain, and strong agreement for effective efficiency, exergy efficiency, and thermal efficiency. The LR model showed superior predictive accuracy, achieving testing R² values of 0.98879 for useful heat gain, 0.97982 for thermal efficiency, 0.97381 for effective efficiency, and 0.93471 for exergy efficiency. This superior performance suggests that the relationships between operating parameters such as solar intensity, flow rate, and inlet temperature and thermal outputs are predominantly linear under the tested conditions, despite the turbulent flow induced by the baffles and PTT inserts. The relatively small dataset (119 samples) and the complexity of the RF and KNN models contribute to mild overfitting, resulting in reduced generalization to unseen data. In practical applications, the simple and interpretable LR model provides sufficient accuracy for predicting the thermal performance of the ETSAH configuration. This enables rapid design optimization and real time performance monitoring without the need for extensive experimental campaigns. The statistical robustness of the LR model is supported by cross-validation and residual diagnostics.</p>