Multidimensional data are commonly indexed using space-filling curves (SFCs), which map data points from a multidimensional space to one-dimensional keys and store them in traditional index structures such as B\(^+\)-trees. However, most existing approaches adopt fixed mapping strategies, which struggle to adapt to complex data distributions and diverse query workloads, leading to degraded performance in high-dimensional or skewed settings. In this paper, we propose HSFC, a learned hybrid space-filling curve indexing framework that combines multi-level spatial partitioning with subspace-level curve optimization. HSFC first constructs a query-aware Multi-level Spatial Partitioning Tree (MSP-Tree) to recursively divide the global data space into localized subspaces. Within each subspace, a monotonic SFC is independently learned via Bayesian optimization, and a globally ordered one-dimensional key space is formed through prefix encoding mechanism from MSP-Tree. To further reduce false positives, we introduce a proactive query splitting strategy leveraging weight pruning that decomposes window queries into multiple subqueries with tighter projection intervals. Extensive experiments on the evaluated real-world and synthetic datasets show that HSFC outperforms the compared baselines, achieving 1.5\(\times \)–8.8\(\times \) reductions in query latency and 27%–88% reductions in false positives.
