A new AI system squeezes a little more material out of circular packing jobs by learning where packing algorithms get stuck.
Researchers built GeoNest, a framework that pairs a reinforcement learning policy with existing geometric packing solvers to tackle the two-dimensional irregular knapsack problem: fitting oddly shaped pieces into a circular container with minimal waste. Standard solvers pack tightly at first but tend to scatter leftover space into small pockets too tiny for any remaining piece, a late-stage bottleneck. GeoNest's graph policy diagnoses which already-placed pieces are blocking a fit, then selects a bounded repair job for the underlying solver to execute. The team tested it on CircleNest-Bench, a new benchmark of 2,391 instances pulled from four contour sources, including one held-out industrial CAD dataset.
Material utilization is a real cost line in manufacturing, since cutting sheet metal, fabric, or composites more efficiently means less scrap per job. GeoNest improved average utilization by about 0.9% on the main test sets and 0.6% on the industrial set, under the same time budget as the best standalone solver it was tested against.
Those are modest numbers, but at manufacturing scale, fractions of a percent compound across thousands of cuts. It's a reminder that AI's most useful manufacturing wins right now are not flashy generative demos but incremental fixes to decades-old optimization problems.