Seasonal forecasts of air-temperature generated by numerical models provide guidance to the planners and to the society as a whole. However, generating accurate seasonal forecasts is challenging mainly due to the stochastic nature of the atmospheric internal variability. Therefore, an array of ensemble members is often used to capture the prediction signals. With large spread in the prediction plumes, it becomes important to employ techniques to reduce the effects of unrealistic members. One such technique is to create a weighted average of the ensemble members of seasonal forecasts. In this study, we applied a machine learning technique, viz. a genetic algorithm, to derive optimum weights for the 24-ensemble members of the coupled general circulation model; the Scale Interaction Experiment-Frontier research center for global change version 2 (SINTEX-F2) boreal summer forecasts. Our analysis showed the technique to have significantly improved the 2m-air temperature anomalies over several regions of South America, North America, Australia and Russia compared to the unweighted ensemble mean. The spatial distribution of air temperature anomalies is improved by the GA technique leading to better representation of anomalies in the predictions. Hence, machine learning techniques could help in improving the regional air temperature forecasts over the mid- and high-latitude regions where the model skills are relatively modest.