How the Classical Baseline Shapes Machine-Learning Speed Claims in Topology Optimization

Key points

• Machine-learning methods often report large speed-ups over classical, non-learned optimization, but a slow classical code or a less stiff learned design can exaggerate them.1
• We re-ran released code for eight methods on one workstation and studied two in depth: Optimize Any Topology (OAT), a diffusion model that generates designs, and a graph-based neural optimizer.1
• Published timings imply a 134x speed-up for OAT on a 256 x 256 grid; our matched single-design comparisons gave 1.5x to 3.5x.1
• At matched stiffness, the graph-based optimizer was slower than the classical code in all six two-dimensional problems.1

Sources (1)

• [1]How the Classical Baseline Shapes Machine-Learning Speed Claims in Topology Optimization
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 01:45 AM

  Topology optimization places material so that a structure resists deformation under given loads.

  Machine-learning methods often report large speed-ups over classical, non-learned optimization, but a slow classical code or a less stiff learned design can exaggerate them.

Extractive summary: sentences quoted from the sources.