A Large-Scale Dataset for Evaluating Encoding, Embedding, and Annealing Strategies for Aircraft Container Loading on D-Wave Systems

This dataset accompanies the paper A Large-Scale Study of Encoding, Embedding, and Annealing Strategies for Aircraft Container Loading on D-Wave Systems. Passenger airlines supplement ticket sales by transporting air cargo. Selecting which cargo to take aboard and arranging it within an aircraft give rise to NP-hard optimisation problems amenable to approximate solution by quantum annealing. The study considers a simplified two-phase aircraft container loading problem: selecting cargo to maximise loaded mass subject to capacity constraints, and arranging the selected containers to minimise the squared moment about the aircraft’s centre. It compares D-Wave quantum annealing with classical solvers across problem encodings, hardware embedding methods, constraint penalty strengths, and forward and warm-started reverse annealing. Quantum experiments were performed on two D-Wave QPUs, Advantage_system6 and Advantage2_system1. The release contains 842,900 experimental records: 215,000 for Phase 1 and 627,900 for Phase 2. Each record corresponds to a problem instance evaluated under a particular solver configuration. The records are provided in two forms. The raw records, one JSON file per record, are in a compressed archive for each phase and include the D-Wave sample sets, unembedded to logical variables by majority vote, and minor embeddings. Two Parquet tables, phase_1.parquet and phase_2.parquet, have one row per record. They include instance parameters, solver settings, penalty and embedding information, timing measurements, feasibility and solution-quality outcomes, and derived metrics describing the encoded problems and embedded chains. Field availability depends on the solver and problem phase. The 163,110 QUBO configurations for which no minor embedding was found were not annealed. For phase 1 embedding outcomes on Advantage_system6, instances 14-99 suffered from a QPU outage which meant that some of these rows failed to anneal. These embeddings were re-embedded to ensure they were not stale in the case of a system calibration and graph change. Since these instances needed to have been successfully embedded to reach this reset, such instances succeeded in being embedded only if the stochastic minorminer embedding routine succeeded on two occasions, with failures at either point resulting in a failed embedding. These instances therefore have lower embedding success rates in the region of problems for which embedding does not typically succeed with 100% probability. The device Advantage2_system1 had 1528 rows reset with only 15 having been previously embedded. All these 15 affected instances for this device embedded successfully on both occasions with, therefore, no impact on the data. Phase 1 instances 0-13 and Phase 2 remain unaffected on both devices.