A novel index structure, \(\text {SPL}^{index}\) , is proposed for large databases of complex geometries. It owes its efficiency to a unique integration of learned indexes, clustering, and dimension reduction by the \(\mathcal {Z}\) -order curve, and is combined with an optimal disk layout. Optimal hyperparameters to control the clustering algorithm can be found by gradient descent, and based on this, a linear and very precise approximation is developed. \(\text {SPL}^{index}\) outperforms the state-of-the-art R-tree for both range and point queries.
