Condensation heat transfer has been widely studied for various applications such as power generation, water desalination, data centers, chemical and pharmaceutical syntheses, heating and air conditioning systems, and especially, the passive containment cooling system (PCCS) in nuclear power plants. The PCCS is designed to reduce the risk resulting from a loss of AC power or compounding human error. In a hypothetical accident situation, one of the main safety systems, the PCCS, condenses water vapor with gravity-driven force to remove the heat from the containment vessel. Although the condensation heat transfer in the presence of non-condensable gases for predicting the heat transfer performance of the PCCS has been successfully investigated, there is still no generalized model or correlation available. In this work, we focused on the development of universal models for predicting the heat transfer coefficients of free-fall condensation heat transfer on the external surfaces of vertical tubes in the presence of non-condensable gases. For this, we created a consolidated database covering a broad range of geometric values and operating conditions, including tube hydraulic diameter Dh = 10–41.2 mm, tube length L = 0.3–3.5 m, total pressure Ptot = 0.15–2.0 MPa, wall subcooling temperature ΔTsub = 5–71 K, average condensation heat transfer coefficients of 90 ≤ ͞h ≤ 19,400 W/m2K, and Reynolds numbers of the film of 8 ≤ Refilm ≤ 7160. We analyzed the influence of varying the geometric and operational parameter values by using this database. Conventional machine learning techniques such as nonlinear regression and multilayer perceptron (MLP) neural network methods were adopted to predict the condensation heat transfer rate based on the consolidated database. Moreover, the prediction accuracies of the condensation heat transfer rates of the proposed prediction models and 12 relevant correlations were compared by using the consolidated database. The proposed nonlinear regression model exhibited good prediction accuracy with a mean absolute error (MAE) of 12.7 % for average ͞h, which is much lower than those achieved by previously proposed relevant correlations. In addition, the MLP neural network model showed excellent prediction accuracy with an MAE of 4.2 % for the consolidated database.