Channel charting is an emerging technology that enables self-supervised pseudo-localization of user equipments by performing dimensionality reduction on large channel-state information (CSI) databases that are passively collected at infrastructure base stations or access points. In this paper, we introduce a new dimensionality reduction method specifically designed for channel charting using a novel split triplet loss, which utilizes physical information available during the CSI acquisition process. In addition, we propose a novel regularizer that exploits the physical concept of inertia, which significantly improves the quality of the learned channel charts. We provide an experimental verification of our methods using synthetic and real-world measured CSI datasets, and we demonstrate that our methods are able to outperform the state-of-the-art in channel charting based on the triplet loss.