Prehistoric pottery, such as Jomon pottery used in the Japanese archipelago from 16,000 to 3000 years ago, reflects the activities of small-scale settled societies and the agricultural era. Recent advances in 3D scanning technology for the analysis and preservation of prehistoric pottery as digital data have proven beneficial for systematic related research. In this study, we proto-typed a multimodal AI model that combines a PointNet and a Transformer decoder to input 3D point cloud data of Jomon pottery and output its caption in Japanese language. We then trained the model for 100,000 epochs using an Adam optimizer with a batch size of 4 and a learning rate of 1e-6, evaluating its accuracy and loss. Although the consistency is insufficient, we were able to develop a prototype capable of generating natural captions in Japanese language from pottery point cloud data. Future developments are expected to produce a high-resolution model capable of recognizing the patterns on Jomon pottery.