Abstract: Outdoor layouts are characterized by the complexity and diversity of boundaries and building shapes, and the non-rectangular contours of traditional villages such as the Huizhou architectural complexes reflect this characteristic, which is less frequently discussed in existing intelligent layout methods. Therefore, a method is proposed to generate the layout of Huizhou architectural complexes within a given irregular boundary. First, each building is represented as a boundary-based vector data form, which accounts for the influence of heterogeneous boundaries on the layout elements; second, a neural network model for building serialization generation is proposed, which is composed of LSTM encoder ang decoder, and the LSTM network is applied to generate the cluster layout by considering the before-and-after formation time between the buildings, which reflects the formation mechanism of the traditional clusters; and finally, a loss function appropriate for the clusters is used to compute the combined loss of the building categories, geometrical data, and intersection and merger ratios, which makes the data generated by network is more in accordance with the genuine layout. According to experimental results conducted on the Huizhou village dataset, the layout evaluation indexes and domain evaluation indexes of the proposed method are superior than those of related layout methods, and the generated Huizhou architectural complex layout is more consistent with the real layout characteristics.