← 返回

面向工业电镀排产的 GNN 加速 MILP 求解

2025年2月15日

Problem

Industrial electroplating lines for aerospace components involve hundreds of thousands of decision variables — tank allocation, job order, drying windows, cross-line synchronisation. Vanilla MILP solves are slow and unstable under operational change.

Contributions

  • Exact formulation of the entire scheduling problem as a large-scale Mixed Integer Linear Program.
  • GNN initial-solution guidance. Train a GNN on historical schedules; at solve time, the model produces (i) a warm-start feasible solution and (ii) a priority ranking over branching variables. Average solve time improves by > 10×.
  • FENNEL streaming partitioning. The bipartite constraint graph (hundreds of thousands of nodes) is split into loosely-coupled subgraphs, cutting GNN train/inference cost by 60–70% and enabling per-subgraph parallel inference.
  • High-confidence variable fixing. The top 10% most confident decisions from the GNN are fixed before optimisation, reducing the effective number of decision variables by ≈ 90% and compressing problem complexity by more than an order of magnitude.