The numerical simulation of concrete materials has been widely accepted, and aggregate modeling based on real materials is crucial. Traditional methods acquire aggregate sequences through techniques like Monte Carlo sampling and arrange them one by one, requiring overlap detection in the process, which is computationally complex and time-consuming. This study introduces a diffusion generative model to learn and reconstruct the distribution characteristics of 2D mesoscopic concrete. It takes full advantage of the similarity in randomness between actual aggregate distribution and the generative model. The actual generation speed is significantly faster than traditional methods. Then, through its application in chloride ion erosion simulation, it is demonstrated that the concrete mesoscopic models generated by the proposed method can meet the requirements for numerical simulation.