Effective grain size prediction for aluminum alloy die castings is of great significance to the rational formulation of die casting process parameters and to the improvement of casting mechanical properties. The traditional grain size prediction method cannot give consideration to both the efficiency and accuracy because of its inherent defects. To improve the efficiency and accuracy of predicting grain size for aluminum alloy die castings, this study proposes a prediction method that is based on the genetic algorithm-extreme learning machine (GA-ELM) model. ELM has the characteristics of few parameter settings, fast learning and good generalization performance, but the algorithm randomly generates the initial input layer weight matrix and the hidden layer threshold matrix, which greatly affects the prediction result. By exploiting GA's excellent global optimization ability, the optimal initial input layer weight matrix and the hidden layer threshold matrix for ELM can be found. The establishment of GA-ELM model can considerably improve the prediction accuracy of ELM model. This study uses grain size as the output parameters and relevant die casting process parameters as the input parameters. The castings produced under different die-casting process parameters are obtained experimentally, and the microstructures of specified sections of key casting positions are analyzed and measured to obtain the average grain size of the sec tions, i.e. the output parameters. The GA-ELM model is trained and tested using these data. To verify the superiority of the GA-ELM model in grain size prediction, this study compares the prediction results of GA-ELM model with the GA-BP neural network model and the original ELM model, and eventually verifies the reliability of GA-ELM model prediction results through metallographic structure measurement experiment. The results show that the GA-ELM model has higher prediction accuracy than the GA-BP neural network model and the original ELM model. Besides, its prediction efficiency is higher than the GA-BP model, while is lower than the original ELM model. With fairly high prediction accuracy and efficiency, the GA-ELM model can meet the actual engineering requirements. Furthermore, its prediction reliability and excellent prediction effect are verified by the results of metallographic structure measurement experiment.