The thickness of thin-walled castings such as aerospace cabin bodies is uneven, which makes it difficult to feed them during solidification. The shrinkage porosity of cylindrical low-pressure aluminium-silicon alloy castings was studied via numerical simulation. A data-driven machine learning program was developed on the basis of the Gaussian regression surrogate model and genetic algorithm. The results show that the shrinkage porosity predicted by machine learning in low-pressure casting is consistent with the numerical simulation results based on the mechanism model. The higher holding pressure with the lower degree of superheating and the shorter filling time lead to a smaller tendency toward porosity formation during casting. Compared with an orthogonal design, machine learning can design reasonable process parameters. The efficiency of predicting porosity via machine learning is greater than that via numerical simulation.