Autoencoder-Driven Clustering of Intersecting D-brane Models via Tadpole Charge

Abstract

We study the well-known type IIA intersecting D-brane models on the $T^6/(\mathbb{Z}_2 \times \mathbb{Z}'_2)$ orientifold via a machine-learning approach. We apply several autoencoder models with and without positional encoding to the D6-brane configurations satisfying certain concrete models described in arXiv:hep-th/0510170 and attempt to extract some features which the configurations possess. We observe that the configurations cluster in two-dimensional latent layers of the autoencoder models and analyze which physical quantities are relevant to the clustering. As a result, it is found that tadpole charges of hidden D6-branes characterize the clustering. We expect that there is another important factor because a checkerboard pattern in two-dimensional latent layers is observed in the clustering.