Dataset for the Bottleneck SINR Optimized Network in Directional mmWave Swarm Topologies

The dataset contains 10,000 labelled instances of the bottleneck-SINR topology optimization problem for multi-hop UAV networks with steerable directional mmWave antennas, split into 8,000 training and 2,000 test samples. Each instance is a snapshot of a swarm placed on a 3D grid with 6 km nominal spacing and zero-mean Gaussian jitter (σ = 1.5 km), following the enhanced Gauss-Markov mobility model at speeds of 100–200 m/s. Links are line-of-sight with Rician fading (K = 15 dB) and the 3GPP TR 38.900 above-6-GHz radiation pattern; transmit power, antenna beamwidth, and minimum degree vary across instances so that models must generalize over operating points rather than a single configuration. Training networks are limited to 32 nodes, while test networks range from 10 to 60, so that extrapolation beyond the training scale can be measured.
Every instance provides the node geometry, the received signal powers and the full interference tensor, the derived line-graph representation with its node and edge features, and the optimal bidirectional link set together with the corresponding threshold τ* and solver runtime. Labels were obtained by solving the MINLP exactly with the bisection-search MILP formulation in SCIP at a precision of 0.05 dB. Because symmetric layouts admit degenerate optima, the released label is one optimal topology rather than the unique one, so achieved bottleneck SINR is the more reliable basis for comparing methods than classification metrics alone.