engiopt-engopt2026-flow-matching-2d-cond

EngiOpt flow_matching_2d_cond

This repository stores evaluation-ready EngiOpt checkpoints for one model family. The models are trained on EngiBench problem datasets and are intended as learned initializations for downstream engineering optimization.

Each package contains model weights, run_config.json, and metadata.json. Selected checkpoint packages also include the validation shortlist and selection results. The metadata links each package to its W&B evaluation run and records the source revisions needed to reproduce the evaluation.

<!-- BEGIN beams2d-sigma1-metric-correction -->

Beams2D metric correction

The selected Beams2D CFM checkpoints were not retrained or reselected. Their final MMD and DPP values were recomputed with the declared Beams2D kernel bandwidth sigma=1.0; the original evaluation used sigma=10.0. COG, FOG, volume-fraction deviation, checkpoint files, and checkpoint checksums are unchanged. The original values remain in each package under test_metrics_pre_correction, and the corrected provenance is recorded in corrections/beams2d_sigma1_v1.json.

• Public W&B report
• Hugging Face collection

<!-- END beams2d-sigma1-metric-correction -->